{
  "id": 265921,
  "title": "Best adversarial model",
  "url": "/competitions/seti-breakthrough-listen/discussion/265921",
  "author_name": "CPMP",
  "post_date": "2021-08-17T11:13:55.803000",
  "votes": 22,
  "comment_count": 3,
  "views": 0,
  "content": "<p>You may be bothered by the best single model threads we see in every comp, so here is a variant: for those who trained an adversarial model what is your best?</p>\n<p>For those not familiar with it, adversarial models are models training on the union of new train data and new test data with target 0 on train and target 1 on test (you can switch targets of course).  You then run a k fold cross validation on shuffled data.</p>\n<p>My 5 fold CV score is 0.994.  It was obtained after data processing that tried to remove some train/test difference.  I didn't dare train without this data processing as this was frightening enough.</p>",
  "messages": [
    {
      "id": 1477135,
      "postDate": "2021-08-17T11:13:55.803Z",
      "content": "<p>You may be bothered by the best single model threads we see in every comp, so here is a variant: for those who trained an adversarial model what is your best?</p>\n<p>For those not familiar with it, adversarial models are models training on the union of new train data and new test data with target 0 on train and target 1 on test (you can switch targets of course).  You then run a k fold cross validation on shuffled data.</p>\n<p>My 5 fold CV score is 0.994.  It was obtained after data processing that tried to remove some train/test difference.  I didn't dare train without this data processing as this was frightening enough.</p>",
      "rawMarkdown": "You may be bothered by the best single model threads we see in every comp, so here is a variant: for those who trained an adversarial model what is your best?\n\nFor those not familiar with it, adversarial models are models training on the union of new train data and new test data with target 0 on train and target 1 on test (you can switch targets of course).  You then run a k fold cross validation on shuffled data.\n\nMy 5 fold CV score is 0.994.  It was obtained after data processing that tried to remove some train/test difference.  I didn't dare train without this data processing as this was frightening enough.",
      "votes": 21
    },
    {
      "id": 1477363,
      "postDate": "2021-08-17T13:00:07.073Z",
      "content": "<p>Shouldn't we be posting worst adversarial model scores 😅</p>",
      "rawMarkdown": "Shouldn't we be posting worst adversarial model scores 😅",
      "votes": 4
    },
    {
      "id": 1477300,
      "postDate": "2021-08-17T12:32:37.570Z",
      "content": "<p>Interesting! My 4-fold CV score is 0.99875.</p>",
      "rawMarkdown": "Interesting! My 4-fold CV score is 0.99875.",
      "votes": 2
    },
    {
      "id": 1477387,
      "postDate": "2021-08-17T13:10:14.293Z",
      "content": "<p>Its worrisome for me as well and led me to post a discussion about it , mine is 0.992  5-fold, trained after some preprocessing as well but it didn't reduce our CV/LB gap </p>",
      "rawMarkdown": "Its worrisome for me as well and led me to post a discussion about it , mine is 0.992  5-fold, trained after some preprocessing as well but it didn't reduce our CV/LB gap ",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 1477363,
      "author_name": "Mohsin hasan",
      "author_url": "",
      "post_date": "2021-08-17T13:00:07.073000",
      "content": "<p>Shouldn't we be posting worst adversarial model scores 😅</p>",
      "votes": 4,
      "replies": []
    },
    {
      "id": 1477300,
      "author_name": "Shion Honda",
      "author_url": "",
      "post_date": "2021-08-17T12:32:37.570000",
      "content": "<p>Interesting! My 4-fold CV score is 0.99875.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1477387,
      "author_name": "Mr_KnowNothing",
      "author_url": "",
      "post_date": "2021-08-17T13:10:14.293000",
      "content": "<p>Its worrisome for me as well and led me to post a discussion about it , mine is 0.992  5-fold, trained after some preprocessing as well but it didn't reduce our CV/LB gap </p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1477135": "You may be bothered by the best single model threads we see in every comp, so here is a variant: for those who trained an adversarial model what is your best?\n\nFor those not familiar with it, adversarial models are models training on the union of new train data and new test data with target 0 on train and target 1 on test (you can switch targets of course).  You then run a k fold cross validation on shuffled data.\n\nMy 5 fold CV score is 0.994.  It was obtained after data processing that tried to remove some train/test difference.  I didn't dare train without this data processing as this was frightening enough.",
    "1477363": "Shouldn't we be posting worst adversarial model scores 😅",
    "1477300": "Interesting! My 4-fold CV score is 0.99875.",
    "1477387": "Its worrisome for me as well and led me to post a discussion about it , mine is 0.992  5-fold, trained after some preprocessing as well but it didn't reduce our CV/LB gap "
  }
}